Disentangling Label Distribution for Long-tailed Visual Recognition (CVPR 2021)

• Arxiv link
• Blog post
• Poster
• This codebase is built on Causal Norm.

Install

conda create -n longtail pip python=3.7 -y
source activate longtail
conda install pytorch torchvision cudatoolkit=10.1 -c pytorch
pip install pyyaml tqdm matplotlib sklearn h5py tensorboard

Training

Preliminaries

• Download pretrained caffe resnet152 model for Places-LT: please refer to link.

• Prepare dataset: CIFAR-100, Places-LT, ImageNet-LT, iNaturalist 2018

  • Please download those datasets following Decoupling.

CIFAR-100 training

For CIFAR-100 with imbalance ratio 0.01, using LADE:

python main.py --seed 1 --cfg config/CIFAR100_LT/lade.yaml --exp_name lade2021/cifar100_imb0.01_lade --cifar_imb_ratio 0.01 --remine_lambda 0.01 --alpha 0.1 --gpu 0

Places-LT training

For PC Softmax:

python main.py --seed 1 --cfg config/Places_LT/ce.yaml --exp_name lade2021/places_pc_softmax --lr 0.05 --gpu 0,1,2,3

For LADE:

python main.py --seed 1 --cfg config/Places_LT/lade.yaml --exp_name lade2021/places_lade --lr 0.05 --remine_lambda 0.1 --alpha 0.005 --gpu 0,1,2,3

ImageNet-LT training

For LADE:

python main.py --seed 1 --cfg config/ImageNet_LT/lade.yaml  --exp_name lade2021/imagenet_lade --lr 0.05 --remine_lambda 0.5 --alpha 0.05 --gpu 0,1,2,3

iNaturalist18 training

For LADE:

python main.py --seed 1 --cfg ./config/iNaturalist18/lade.yaml --exp_name lade2021/inat_lade --lr 0.1 --alpha 0.05 --gpu 0,1,2,3

Evaluate on shifted test set & Confidence calibration

For Imagenet (Section 4.3, 4.4):

./notebooks/imagenet-shift-calib.ipynb

For CIFAR-100 (Supplementary material):

./notebooks/cifar100-shift-calib.ipynb

License

The use of this software is released under BSD-3.

Citation

If you find our paper or this project helps your research, please kindly consider citing our paper in your publications.

@inproceedings{hong2021disentangling,
  title={Disentangling label distribution for long-tailed visual recognition},
  author={Hong, Youngkyu and Han, Seungju and Choi, Kwanghee and Seo, Seokjun and Kim, Beomsu and Chang, Buru},
  booktitle={Proceedings of the IEEE/CVF conference on computer vision and pattern recognition},
  pages={6626--6636},
  year={2021}
}